CommentaryIn this five-part series, data analyst David Stasior argues that AI has made it possible to turn court and disciplinary records into data at scale, allowing the rule of law to be measured rather than just assumed. Part 1 made the case, and Part 2 put it to work on 24 years of Massachusetts attorney disciplinary decisions. Here, in Part 3, he turns to federal court statistics, where AI can quickly uncover insights linked to the rule of law, despite awkward formatting that makes the numbers hard to use.
For the rule of law, AI lets us test deeply held beliefs against empirical reality. In Part 2, we used it to turn thousands of unstructured state disciplinary records into a structured dataset. The analysis uncovered a disciplinary authority operating more like a legal consumer-complaints bureau than an arbiter of ethical misconduct. Dishonesty, in particular, was treated with striking variability. Sometimes a lie drew no additional discipline at all, but sometimes it led to an admonition, a public reprimand, or a suspension of months or years — despite precedent calling for at least a one-year suspension for a knowingly false statement.
Part 2 closed with a choice: cling to our comfortable assumptions, lower our expectations, or change how the profession is governed. But AI applies to more than attorney discipline. The same tools can test our beliefs about the federal courts and how they uphold the rule of law. The results, of course, could leave us facing the same three choices.
For the federal courts, however, the data landscape changes drastically. We are confronted with a flawed source: court administrative data published in oddly constructed Excel workbooks and PDF tables, with no single bulk download available to the public.
The Administrative Office of the US Courts posts its files in an online directory. In October 2026, an AI agent found 168 directory pages listing more than 10,000 files, about two-thirds PDFs and one-third Excel workbooks. Most are new editions of the same tables; the agent counted 247 distinct table types, nearly all reissued on a regular schedule.
The Administrative Office essentially treats Excel as a typesetting tool rather than a database. It publishes visual “reports” that happen to be saved as .xlsx files, with merged cells, multi-level headers, and footnotes scattered through the cells. The Excel files look almost identical to the PDF versions. This structure means even a simple sort is nearly impossible. And each table stands alone, with few common fields to link it to the next.
During the first week of October 2026, I directed an AI agent to download all the files. It retrieved all 10,353 in 22 minutes and 35 seconds. Within a few minutes more, the agent “read” the tables, reformatted the data, and completed its first analysis. Over the next hour, with additional prompts, I ran several more rounds of analysis.
At the end of Part 2, I asked what share of cases underlies the famous “rocket docket” reputation of the US District Court for the Eastern District of Virginia. The court measures its speed by the time from filing to the start of a civil trial, and by that measure it ranked first in the country in 2025. By a longer-standing measure, the time through the end of trial, it ranked second.
Now ask yourself: what percentage of the district’s civil cases are actually used to compute the speed that defines the “rocket docket”? If you wrote down a prediction after Part 2, now is the time to check it.
When we go to the source, the court’s official tables, the arithmetic is simple and the reality is jarring. Of the 2,960 civil cases the district resolved in 2025, the longer-standing speed ranking, measured through the end of trial, rested on exactly 16 trials. That is roughly 0.54 percent of those cases. The district’s own preferred measure, time to the start of trial, is no more reassuring: it rests on 17 trials.
Bestowing a reverent title based on half of 1 percent of a court’s cases is odd. It is like declaring an NFL team the fastest in the league based solely on how quickly a 300-pound lineman runs after recovering a fumble. Even if every team is measured similarly with this single metric, the ranking ignores the running backs, wide receivers, tight ends, and quarterbacks who carry nearly every play.
Now imagine telling the head coach of the “fastest” team how the title was awarded. He would dismiss it as trivial. Anywhere in the league, the measure would not survive the “smell test.”
In the federal justice system, the “rocket docket” title is based on a similarly flimsy statistic. Yet each year, a prominent litigator publishes an analysis of the district’s speed. The 2026 edition even reports how rarely civil cases go to trial but never connects that paltry percentage to the silliness of this speed title. Instead, in an interview, the district’s chief judge credits the hard work and commitment of the court and the bar. Few of the court’s staff would have touched those trials. In the legal community, belief seems to outrank reality.
Perhaps, though, the moniker is justified for other reasons. The court’s tables hold better measures of speed than these trial counts. The most basic is the median time to resolve a civil case of any kind. By that measure, the Eastern District of Virginia is genuinely fast: 5.6 months in the 12 months ending September 30, 2025, quicker than about 90 percent of districts. But the tables also show what remains unresolved. About one in six of its pending civil cases, 549 of 3,322, had been open at least two years, a share near the middle of all districts.
I asked AI to combine these and other measures into composite rankings, weighting them in different ways. In the civil-only version, Southern Florida ranked first and the Eastern District of Virginia ranked 16th. In this and other versions, the district routinely landed in the top 20 but never ranked first. With simple prompts for AI, I could readily assess rankings based on a variety of models.
Adding criminal cases reshuffled the list. Hawaii fell from 8th to 49th, and Central California from 3rd to 33rd. Western Texas rose from 37th to 8th, and Southern Texas from 35th to 7th. Districts along the southwest border close felony cases in about four to five months, compared with 16 months in Central California and 21 in Hawaii, likely reflecting fast-moving immigration prosecutions. The Eastern District of Virginia rose to 10th. Which court is “fastest” depends heavily on what is measured.
This exercise showed that AI can make quick work of data organized for stiff reports rather than analysis. In less than an hour, a reputation as the fastest district became a set of measurable claims, confirming the overall speed of the Eastern District of Virginia but passing the title of fastest to a fellow district. Tables built for printed reports became data that can be sorted, joined, and compared.
The real value from AI is not criticizing the precision of a long-standing moniker. Instead, each district could use this data to understand its own operations and identify how it could deliver justice more effectively. The Eastern District of Virginia resolves typical civil cases quickly. But it ranks near the middle on civil cases left open for years. It also ranks near the middle on the time from conviction to sentencing, 117 days.
These operations directly impact the rule of law. One key feature of the rule of law is a legal system that is predictable, clear, and accessible to all. How does the variability in timing affect all parties involved in federal cases? The district’s chief judge herself invokes the adage “justice delayed is justice denied.” The tables let us measure related variables: how long people wait for a ruling, a trial, or a sentence, and whether that wait depends on where they happen to file or be charged. Enormous variability suggests someone is suffering from a system that does not ensure predictable, clear outcomes. This data is public, spread across thousands of spreadsheets, and waiting for just this kind of thorough analysis. The opportunity to strengthen the rule of law is sitting in plain sight.
These statistics can take us only part of the way. Perhaps differences in court “speed” reflect the type of underlying case, and case types vary by geography. Or perhaps a driver is the rate of plea agreements or the influence of certain prosecutors. To understand the rule of law accurately, we need access to the case records themselves. For the attorney discipline analysis, we relied on exactly this level of detail.
For the federal courts, these records sit behind the paywall of PACER, the federal judiciary’s fee-based online records system. With bulk access, we could explore what drives the differences. For the Eastern District of Virginia, AI could read the dockets of the 17 trials behind its first-place ranking and determine whether they were exceptionally complex antitrust matters or simple contract disputes, then compare them to similar trials in other districts. But because the system shields its documents, we are left with a statistical fragment: enough to show that something varies, not enough to say why.
A full understanding of how the courts uphold the rule of law depends on both the court statistics and the details in the case records. The statistics show where outcomes differ; the documents uncover the reasons. AI makes both practical at scale. The statistics are public, if awkward. The documents are not.
Some of these documents do reach the public, and AI can help us understand how those decisions are made.
Consider a federal defendant on supervised release, the court-ordered monitoring that follows a prison sentence. More than 100,000 people are on supervised release at any given time, for a median term of about three years. After the first year, they can ask the court to end their supervision early, saving the Probation Office the cost of monitoring.
But what would justify ending it early? Imagine a defendant who completed an intensive reentry program, was so successful at work that the store promoted him to manager, stayed sober, and was on the lowest level of supervision. He asked the judge to end his supervision 73 days early. The government objected.
Is staying out of trouble, and even transforming his life, reason enough to end supervision early? Or is that simply what supervised release expects, not grounds for ending it early? If you were the judge, how would you decide, and what reason would you give?
What the judge actually did comes in the next article.
David Stasior, MD, MPP, is a principal at Broken Scales, where he applies data analysis to the justice system, and the author of Slot Machine Justice.
